Source Club — Case Study Submission

My response to the Source Club case study for the Head of AI Powered Operations, Systems & RevOps role. One working app (Assignment 1) plus concise decision docs (2 & 3), a self-narrating video walkthrough (4), and a bonus 90-day architecture scope.

Reviewers — start here: INTERVIEWER_GUIDE.md is the 2-minute "how to run it and what to look at" guide. Everything below links straight to each deliverable.

---

Deliverables — at a glance

#	Assignment	What it is	Open it
1	Savings Analysis Automation	Working app + 3-pass matching pipeline + sample data	code & architecture · run it ↓
2	Stripe × HubSpot Integration	Decision doc: 3 options → recommendation (n8n), schema, multi-location model	read
3	Project Prioritization	The real project queue, ranked: top 5 for the first 90 days + what I'd defer	read
4	Video Walkthrough	Self-narrating demo video (built-in voiceover) + recording script	folder · video ↓
⭐	90-Day Architecture Scope (bonus)	Platform decision + end-state system design + cost model	read

---

Savings Analysis Pipeline — Architecture

flowchart TD
    subgraph INPUT["Data Ingestion"]
        A1["📄 Purchase History CSV\n(buyer spend records)"]
        A2["📋 Supplier Catalog CSV\n(vendor items + pricing)"]
    end

    subgraph PIPELINE["3-Pass Matching Engine"]
        B1["Pass 1 — Exact Match\nSKU / product code lookup"]
        B2["Pass 2 — Fuzzy Match\nRapidFuzz token sort ratio\n≥ 85% threshold"]
        B3["Pass 3 — Claude AI Pass\nSemantic understanding\nAmbiguous line items"]
    end

    subgraph OUTPUT["Analysis & Reporting"]
        C1["📊 Savings Summary\n~$4,944 identified\n78.6% match rate"]
        C2["🔍 Match Confidence\nHigh / Medium / Low tiers"]
        C3["📥 CSV Export\nAll matched line items"]
    end

    A1 --> B1
    A2 --> B1
    B1 -->|"Unmatched"| B2
    B2 -->|"Still unmatched"| B3
    B1 -->|"19 high-confidence"| C1
    B2 --> C1
    B3 --> C1
    C1 --> C2
    C2 --> C3

Loading

Result on sample data: ~$4,944 in savings at 78.6% match rate (19 high-confidence matches).
Works with no API key (fuzzy-only); add ANTHROPIC_API_KEY in .env to enable Claude pass.

---

Run the app (Assignment 1)

pip install -r requirements.txt
streamlit run app.py            # opens http://localhost:8501

On the Savings Analysis page: click Load sample on both uploaders → Run.

Optional one-click hosting (Render / Hugging Face) in docs/DEPLOY.md.

---

Watch the demo (Assignment 4) — with audio

A finished, self-narrating walkthrough covering all three assignments — the live savings tool (1), the Stripe×HubSpot recommendation (2), and the prioritization (3) — with a natural neural voiceover, ~3 min, no recording required. Three ways to play it, all with sound:

• On GitHub (no install): open demo/output/source-club-demo-narrated.mp4 → GitHub plays the MP4 inline. Press play, sound on.
• Locally: open demo/output/source-club-demo-narrated.mp4 in any browser/player.
• In the app: streamlit run app.py → Video Walkthrough page → press play.

---

Integration Architecture — Stripe × HubSpot (Assignment 2)

flowchart LR
    subgraph SOURCES["Revenue Sources"]
        S1["💳 Stripe\nPayments + Subscriptions"]
        S2["🏢 HubSpot CRM\nDeals + Contacts"]
        S3["📍 Multi-Location\nPOS / In-store"]
    end

    subgraph INTEGRATION["n8n Integration Layer"]
        N1["Stripe Webhook Trigger"]
        N2["HubSpot Sync"]
        N3["Dedup + Merge Logic"]
    end

    subgraph UNIFIED["Unified Revenue View"]
        U1["Single Customer Record"]
        U2["Revenue Attribution"]
        U3["Churn Signals"]
    end

    S1 --> N1
    S2 --> N2
    S3 --> N3
    N1 --> N3
    N2 --> N3
    N3 --> U1
    U1 --> U2
    U1 --> U3

Loading

---

How I approached it

• Assignment 1 is running code you can try right now, structured as the team's own two-part process (collect & clean the purchase history → run the analysis).
• Assignments 2 & 3 are concise, act-on-able decision docs — Assignment 3 prioritizes the actual project queue, not invented projects.
• The architecture reflects what I'd actually build, with explicit notes on what the POC skips.

Confirmed vs. assumed: the brief confirms only Stripe and HubSpot. Google Workspace, GCP, ZenOne, Base86, and PandaDoc are working assumptions — flagged throughout.

---

Repository layout

app.py                         Multipage Streamlit app — run this
INTERVIEWER_GUIDE.md           Reviewer quick-start (2 min read)
pages/                         5 pages: savings tool + 4 doc pages
assignment-1-savings-analysis/ matcher.py · report_generator.py · sample_data/
assignment-2-stripe-hubspot/   Decision doc
assignment-3-prioritization/   Real-queue prioritization
assignment-4-video/            Demo script — video in demo/output/
demo/                          record_narrated.py · voiceover-script.md
docs/                          Case study brief · 90-day architecture · DEPLOY.md
.github/workflows/             CI (test) + Deploy (Azure Container Apps)

---

Supporting docs: docs/CASE-STUDY-BRIEF.md · docs/questions-i-would-ask-first.md · docs/DEPLOY.md